MaxPool2d

MaxPool2d reduces image size and preserves important features.

MaxPool2d

Pooling shrinks the feature map. nn.MaxPool2d(2) takes the largest value from every 2×2 region: height and width are halved, the channel count doesn't change. It has no parameters:

havuz = nn.MaxPool2d(2)
x = torch.tensor([[[[1., 3., 2., 4.],
                    [5., 6., 1., 2.],
                    [7., 2., 9., 0.],
                    [1., 8., 3., 4.]]]])
print(havuz(x).tolist())   # [[[[6.0, 4.0], [8.0, 9.0]]]]

Pooling gives robustness to small shifts and reduces the computation of the following layers.

Average and global pooling

nn.AvgPool2d takes the region's mean. Modern networks usually use global average pooling at the end: it reduces each channel to a single number, so the model becomes independent of the input size:

gap = nn.AdaptiveAvgPool2d(1)
print(gap(torch.zeros(8, 512, 7, 7)).shape)    # torch.Size([8, 512, 1, 1])
print(gap(torch.zeros(8, 512, 13, 9)).shape)   # still (8, 512, 1, 1)

AdaptiveAvgPool2d takes the target output size and works out the needed window itself.

Exercise Pooling

Apply nn.MaxPool2d(2) and nn.AvgPool2d(2) to the given 4×4 image; print the two results with tolist() on separate lines.

Show example solution

import torch
from torch import nn

x = torch.tensor([[[[1., 3., 2., 4.], [5., 6., 1., 2.], [7., 2., 9., 0.], [1., 8., 3., 4.]]]])

print(nn.MaxPool2d(2)(x).tolist())
print(nn.AvgPool2d(2)(x).tolist())

Test yourself

What is the shape after applying <code>MaxPool2d(2)</code> to a tensor of shape (8, 32, 16, 16)?
- (8, 16, 16, 16)
- (4, 32, 16, 16)
- (8, 32, 14, 14)
- (8, 32, 8, 8)
Show answer

(8, 32, 8, 8) — Only the spatial dimensions are halved.

Apply 2×2 max pooling.

havuz = nn.____(2)

Show answer

havuz = nn.MaxPool2d(2)

Match the poolings.

MaxPool2d(2)

Region's maximum
AvgPool2d(2)

Region's mean
AdaptiveAvgPool2d(1)

One value per channel
parameter count

0

Next lesson: Feature map

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies